PILOT

Project retained features into a 2D or 3D instance space

Syntax

from instancespace.stages.pilot import PilotStage

outputs = space.run_until_stage(PilotStage)

Description

Standard PILOT learns a linear projection while reconstructing features and algorithm performance. cost_weight controls the performance-reconstruction block. Use an analytic solution when requested and viable; otherwise the numerical path evaluates multiple starts and selects by distance preservation.

The PLS alternative uses SIMPLS. It stores the fitted feature mean and applies (X - x_mean) @ A.T to new instances. Standard projections apply X @ A.T without that PLS mean subtraction.

Parameters

method is standard or pls; dims is 2 or 3. analytic is used only for standard projection. n_tries controls generated numerical starts. A supplied x0 matrix provides one start per column. A valid precalc_alpha bypasses the numerical solve.

For p features, q algorithm columns and dimension d, numerical solution vectors have d * (2*p + q) values. See PILOT options for all fields and interactions.

Viewpoints and Rotation

Three-dimensional projections store an optimized global camera view or one view per view_groups entry. Groups use zero-based algorithm indices. adjust_rotation=True is a 2D rigid rotation for consistent visual orientation; it cannot be used with 3D.

Seeds control starts, not numerical-library arithmetic. Optimizer trajectories and restart winners can differ across environments. Do not rotate coordinates independently of the fitted model's classifiers or geometry.

Returns

Projection matrix, coordinates, reconstruction parameters and error, fitted PLS mean when applicable, numerical restart diagnostics and viewpoints. See the PILOT API.

Examples

from dataclasses import replace
from instancespace.data.options import PilotOptions

projection_options = replace(options, pilot=PilotOptions.default(
    method="pls", dims=3,
))